Research On Fluorescence Lifetime Imaging Data Analysis Method Based On Deep Learning

Cancer is a serious threat to human life and health.The accurate identification of early gastric mucosal cancer is of great significance to improve the accuracy of disease diagnosis and reduce the morbidity and mortality of gastric cancer^[1].Fluorescence lifetime microscopic imaging（FLIM）technology achieves quantitative characterization of morphological structural and metabolic differences between normal and tumor cells,which poses a promising future for clinical applications in tumor detection.However,FLIM data analysis has defects of data processing and low efficiency.Relying on the National Natural Science Foundation of China project,this paper proposes a FLIM data processing method based on deep learning network to address the problems in data processing of FLIM technology.We use 3D convolutional neural network（3D-CNN）to efficiently train the simulated data set.High performance reconstruction of fluorescence lifetime parameters is achieved at extremely low photon counting.We have achieved fast and efficient FLIM data analysis and made an experiment on FLIM data of different cancerous periods of gastric mucosa.The main research contents include:1)In term of FLIM imaging,we have took the existing two-photon microscopy imaging（TPM-FLIM）system to complete the data acquisition of human gastric cancer.Due to the problems of low photon number and low signal-to-noise ratio of TPM-FLIM imaging,the Binning technology is used to improve the fluorescence lifetime accuracy of gastric cancer FLIM data.The FLIM data pre-processing algorithm has been developed,and the interface of FLIM data with neural network applications are under study.2)The study of FLIM analysis network model is based on 3D-CNN.By using 3D-CNN structure,we have took the fluorescence（x,y,t）as the decay 3D data input,which presents a scheme about predicting fluorescence parameters such as short lifetime（?₁）,long lifetime（?₂）and fractional amplitude（_RA）of short lifetime from image data.The performance of this network model is evaluated by using t-SNE,SSIM algorithm,mean absolute error（MAE）at different signal-to-noise ratios,etc.The comparative analysis between the performance of network model and least square fitting（LSF）method is completed.3)In order to obtain the massive data sets needed for training,to ensure the robustness of the network structure and the efficiency of feature extraction,and to determine its quantitative accuracy,the MNIST handwritten digital data set is used to study the three-dimensional time point spread function（TPSF）,which realizes the simulation of 3D fluorescence attenuation information on individual pixels.4)We have used a 3D-CNN-based FLIM analysis network to analyze the FLIM data of different cancerous stages of human gastric mucosa,including normal,chronic gastritis with erosion（CG-E）,chronic gastritis with intestinal metaplasia（CG-IM）,and intestinal-type adenocarcinoma（ITA）.The practical value of this network in the field of gastric cancer identification is verified.